The standard scorecard for personal loan origination in Vietnam and Indonesia was designed around a salaried borrower: monthly payroll deposit on a predictable date, employment history verifiable through a tax ID or social security record, debt-to-income ratio calculable from declared income and existing credit obligations. That model works well for about half the urban workforce. For the other half, the model systematically underestimates creditworthiness.
Platform gig workers are the clearest example. A Grab driver who has been on the platform for three years, drives 5 to 6 days per week, receives weekly Grab payouts directly to their GrabPay wallet, and has never missed an informal payment obligation in the transaction history we can see is, by most reasonable definitions of creditworthiness, a better risk than a recent formal-sector hire with two months of payroll history and a thin bureau file from a single credit card. The salaried hire scores higher on a standard model because the model was built to recognize salary deposits. The gig worker's income looks irregular to a feature set that expects bimonthly payroll cycles.
This isn't a fringe problem. Platform economy workers now represent a substantial portion of the urban labor force in Vietnam, Indonesia, and the Philippines. Gojek, Grab, and their regional variants have created a new income category that existing financial infrastructure wasn't designed to evaluate. Addressing this requires rethinking which features carry signal for this population, not just extending existing models to a new segment.
Why Standard Income Features Misread Gig Workers
Standard income verification and feature extraction treats income as a monthly flow. The typical credit model computes average monthly income over the past 3 to 6 months, measures variance around that average, and applies a penalty to borrowers whose income variance exceeds a threshold. The implicit assumption is that variance indicates instability.
For a salaried worker, this assumption is mostly correct. Variance in a salary deposit stream usually means job changes, gaps in employment, or income disruption. For a gig worker, variance in weekly inflows is structural: gig income correlates with hours worked, which varies week to week based on driver or rider choice. A gig worker who takes two weeks off to visit family over Tet will show a dip in inflows during that period and a recovery after return. That pattern looks like income instability in a monthly variance model. It looks like expected behavior in a model designed for platform workers.
The second systematic error is in regularity detection. Monthly variance measures don't capture the higher-frequency regularity that characterizes consistent gig workers. A driver who works every week, with weekly inflows that vary in amount but almost never have a zero-inflow week, is showing a different behavioral signal from someone whose inflows are sporadic and gapped. Rolling weekly inflow frequency, measured over a 26-week window, distinguishes these two populations clearly. The consistent worker has near-zero zero-inflow weeks; the episodic worker has frequent gaps. Loan repayment behavior correlates with this distinction.
Income-Regularity Features That Actually Discriminate
We've iterated through several feature constructions for gig worker income across multiple cohorts. The ones that consistently provide lift for this population share a common structure: they measure regularity of the income pattern rather than magnitude of the income flow.
Zero-inflow week frequency. Over a rolling 26-week window, the fraction of weeks with zero platform inflow. This is a simple binary per-week measure that captures work continuity. A borrower with a zero-inflow frequency above 20 to 25 percent is showing episodic platform engagement, which correlates with weaker repayment behavior. This feature adds more discriminating power for gig populations than standard income variance.
Coefficient of variation with outlier exclusion. Standard coefficient of variation on weekly inflows is sensitive to outlier weeks, particularly high-inflow weeks during peak demand periods (holidays, events). Winsorizing the top 5 percent of weekly inflows before computing CV removes this sensitivity and gives a more stable measure of day-to-day income variation for the typical working week.
Income trend slope. A linear trend fitted to monthly inflow totals over 12 months detects whether the borrower's platform income is growing, stable, or declining. Growing platform income indicates increasing engagement and skill development on the platform. Declining income may indicate health issues, platform deactivation risk, or transition to lower-margin work. The slope is more predictive than the level for medium-term credit outcomes.
Platform payout concentration. For borrowers receiving income from multiple platforms (some drivers work Grab and Gojek simultaneously, some food delivery workers use multiple platforms), the concentration of income across platforms matters. A worker heavily concentrated in a single platform has both more predictable income timing and higher income volatility risk from platform-specific events. This is a feature that needs to be designed with the specific market in mind, since platform concentration patterns differ significantly between Vietnam, Indonesia, and the Philippines.
A Concrete Scenario: Reunderwriting a Declined Application
Consider an application profile we worked with during model development in early 2025: a ride-hail driver in Ho Chi Minh City, 28 years old, three years on the Grab platform. Bureau file: one credit card opened 14 months ago, 10 months of on-time payments, thin by any measure. Standard bureau-primary model score: borderline decline. Monthly income declared on application: approximately 12 million VND.
Running the same application through a cash-flow model with the gig-worker features above produced a different picture. Zero-inflow week frequency over the prior 26 weeks: 3 percent (one week with near-zero inflow, likely a holiday week). Income trend slope over 12 months: modestly positive. Inflow regularity coefficient of variation (winsorized): 0.31, indicating moderate but not high variance. Minimum balance behavior: consistently maintained a balance of approximately 5 to 8 percent of average monthly inflow between payout cycles. Outflow-to-inflow ratio: stable across the 12-month observation window.
The cash-flow model scored this applicant considerably above the bureau model's output, placing them in a tier that, in the validation cohort, had realized default rates consistent with the model's prediction. The bureau model had essentially discounted most of the relevant behavioral evidence because it wasn't designed to process the patterns it contained.
We want to be precise about the boundary here: we're not saying bureau data is irrelevant for this population. The bureau model's low score reflected a short and thin credit history, which is genuine information about repayment track record. What we're saying is that a short bureau history combined with strong behavioral cash-flow signals should produce a different credit assessment than a short bureau history with no supplementary behavioral evidence. Collapsing both into the same decline decision is a modeling error, not a conservative risk policy.
Designing the Training Cohort Correctly
Building accurate models for gig worker populations requires careful attention to training cohort construction. If your historical origination population skewed heavily toward salaried borrowers (because that's who you historically approved), a model trained on that population will inherit the features that discriminated within it and will have limited exposure to what gig worker behavior actually looks like in a performing loan.
The minimum requirement is a gig-worker subsegment in the training data with sufficient positive and negative outcomes to calibrate features properly. In practice this means either a separate model for this population or a mixed model with sufficient representation and explicit gig-worker indicator features. We've found that a hybrid approach, using a shared feature set with a gig-worker segment indicator and segment-specific feature weight calibrations, outperforms both a fully separate model (insufficient data in early training stages) and a fully pooled model (dilutes gig-specific features).
The validation holdout population needs the same care. Performance metrics for the overall model can look acceptable while performance specifically on the gig segment is poor, if the segment is a small minority of the validation set. Segment-level Gini and KS statistics are necessary, not just portfolio-level numbers.
What This Means for Platform Partnerships
The cleanest version of gig worker credit scoring comes from direct data partnerships with the platforms themselves. When a lender has a partnership with Grab or Gojek that provides structured payout history and earnings data directly from the platform API, the data quality and coverage are substantially better than inferring the same information from wallet transaction history.
Platform-direct data eliminates the inference step: you know that a specific inflow is a Grab payout rather than inferring it from a transfer amount and timing pattern. You can get activity metrics that aren't visible in wallet data, like completed ride counts, service rating trends, and account standing signals. You also get data on earnings for periods where the driver may not have transferred their wallet balance to an external account, which wallet-only data misses.
Platform partnerships take time to negotiate and require the platform to have an incentive to cooperate, which usually means the lender is offering something in return, often preferential or embedded credit products for the platform's driver base. But the data quality premium is real, and building these relationships early positions a lender well as gig employment continues to grow as a share of the urban labor market in each of these markets.
